Sparse point‐voxel aggregation network for efficient point cloud semantic segmentation

Effective and efficient semantic segmentation of 3D point cloud data is important for many tasks. Many methods rely on computationally expensive sampling grouping layers to process irregular points, while others convert points into regular volumetric grids them with a U-Net-based network. However, most these suffer from high computational costs cannot be applied the real-time processing large-scale clouds. To address issues, we propose point-voxel-based network architecture named Sparse Point-Voxel Aggregation Network (SPVAN) segmentation. It consists an encoding layer that sparse convolution MLP new decoding called Point Feature Layer (PFAL) only composed feature interpolation layers. Compared recent popular network, our method does not need networks in thus achieves higher speed. Experimental results SemanticKITTI dataset show gets good balance between efficiency performance. Moreover, on-par or better performance than previous challenging S3DIS dataset.